getdoc4313. 279KB Jun 04 2011 12:04:19 AM

J
i
on
Electr

o

u

a
rn l

o

f
P

c

r

ob
abil
ity

Vol. 11 (2006), Paper no. 28, pages 700–722.
Journal URL
http://www.math.washington.edu/~ejpecp/

High-resolution quantization and entropy coding for
fractional Brownian motion
S. Dereich and M. Scheutzow
Institut f¨
ur Mathematik, TU-Berlin
Straße des 17. Juni 136
D-10623 Berlin
{dereich,ms}@math.tu-berlin.de
Abstract
We establish the precise asymptotics of the quantization and entropy coding errors for fractional Brownian motion with respect to the supremum norm and Lp [0, 1]-norm distortions.
We show that all moments in the quantization problem lead to the same asymptotics. Using
a general principle, we conclude that entropy coding and quantization coincide asymptotically. Under supremum-norm distortion, our proof uses an explicit construction of efficient

codebooks based on a particular entropy constrained coding scheme.
Key words: High-resolution quantization; complexity; stochastic process; entropy; distortion rate function
AMS 2000 Subject Classification: Primary 60G35, 41A25, 94A29.
Submitted to EJP on March 22 2005, final version accepted June 1 2006.

700

1

Introduction

Functional quantization and entropy coding concern the identification of “good” discrete approximations to a non-discrete random signal (original ) in a Banach space of functions. These
approximations are required to satisfy a range constraint in the context of quantization and an
entropy constraint in the context of entropy coding. Such discretization problems arise naturally
when digitizing analog signals in order to allow storage on a computer or transmission over a
channel with finite capacity.
As another application, the approximating functions of good quantizers may serve as evaluation
points of quasi Monte Carlo methods. Moreover, some Monte Carlo methods use appropriate
quantizers to carry out a variance reduction (see for instance (18) and the references therein, or
(10)).

Previous research addressed, for instance, the problem of constructing good approximation
schemes, the evaluation of the theoretically best approximation under an information constraint,
existence of optimal quantizers and regularity properties of the paths of optimal approximations.
For Gaussian measures in Hilbert spaces optimal quantizers exist and all its approximating
functions are elements of the reproducing kernel Hilbert space (16). Under mild assumptions
the best achievable distortion in both problems (the quantization and entropy coding problem)
coincide asymptotically and do not depend on the moment under consideration ((17), (4)).
Moreover, this (optimal) approximation error is asymptotically equivalent to the distortion rate
function which can be expressed implicitly in terms of the eigenvalues of the covariance operator.
When the underlying space is a Banach space, the approximation errors of both problems are
weakly asymptotically equivalent to the inverse of the small ball function in many cases ((8),
(5)). Thus asymptotic estimates for the small ball function can be translated into asymptotic
estimates for the above coding problems (see for instance (15) for a summary of results on small
ball probabilitites). Moreover, many approximation quantities of Gaussian measures are tightly
connected to the quantization numbers (see (14), (2)). See also (12) for existence and pathwise
regularity results of optimal quantizers.
The above questions are treated for Gaussian measures in Hilbert spaces by Luschgy and Pag`es
((16), (17)) and by the first-named author in (4). For Gaussian originals in Banach spaces, these
problems have been addressed by the authors and collaborators in (8), (9), (4), (5) and by Graf,
Luschgy and Pag`es in (12). For general accounts on quantization and coding theory in finite

dimensional spaces, see (11) and (1) (see also (13)).
In this article, we consider the asymptotic coding problem for fractional Brownian motion under
supremum and Lp [0, 1]-norm distortion. We derive the asymptotic quality of optimal approximations. In particular, we show that efficient entropy constrained quantizers can be used to
construct close to optimal high resolution quantizers when considering the supremum norm.
Moreover, for all of the above norm-based distortions, all moments and both information constraints lead to the same asymptotic approximation error. In particular, quantization is asymptotically just as efficient as entropy coding. The main impetus to the present work was provided
by the necessity to understand the coding complexity of Brownian motion in order to solve the
quantization (resp. entropy constrained coding) problem for diffusions (see (7) and (6)).
Let (Ω, A, P) be a probability space, let H ∈ (0, 1) and let X = (Xt )t≥0 denote fractional
Brownian motion with Hurst index H on (Ω, A, P), i.e. (Xt )t≥0 is a centered continuous Gaussian
701

process with covariance kernel
1
K(t, u) = [t2H + u2H − |t − u|2H ],
2

t, u ≥ 0.

For a > 0, let C[0, a] and D[0, a] denote the space of continuous real-valued functions on the
interval [0, a] and the space of c`adl`

ag functions on [0, a], respectively. Moreover, we let (Lp [0, a], k·
p
kLp [0,a] ) denote the standard L -space of real-valued functions defined on [0, a]. Finally, k · ks ,
s ∈ (0, ∞], denotes the Ls -norm induced by the probability measure P on the set of real-valued
random variables.
Let us briefly introduce the concepts of quantization and entropy coding. For fixed a > 0 let
d : C[0, a] × D[0, a] → [0, ∞) be a measurable function For a C[0, a]-valued r.v. Y (original ) and
moment s > 0, the aim is to minimize


d(Y, π(Y ))
(1)
s
over all measurable functions π : C[0, a] → D[0, a] with discrete image (strategies) that satisfy a
particular information constraint parameterized by the rate r ≥ 0.

Often we associate a sequence of probability weights (pw )w∈im(π) to a strategy π. Then due
to Kraft’s inequality, there exists a prefix-free representation for im(π) which needs less than
(− log2 pw ) + 1 bits to represent w ∈ im(π). Thus the pair (π, (pw )) corresponds to a coding
scheme translating the original symbol x into a prefix-free representation for π(x). The best

average code length is achieved for pw = P(π(Y ) = w), which leads to an average code length of
about H(π(Y ))/ log 2 (see for instance (1), Theorem 5.2.1).
Entropy coding (also known as entropy constrained quantization in the literature) concerns the
minimization of (1) over all strategies π having entropy H(π(Y )) at most r. Recall that the
(Z)
entropy of a discrete r.v. Z with probability weights (pw ) is defined as
X
(Z)
(Z)
H(Z) = −
p(Z)
w log pw = E[− log pZ ].
w

The entropy constraint represents an average case complexity constraint.
In the quantization problem, one is considering strategies π satisfying the range constraint:
| range (π(Y ))| ≤ er which is a static complexity constraint. The corresponding approximation
quantities are the entropy-constrained quantization error

(2)
D(e) (r|Y, d, s) := inf
d(Y, π(Y ))
s ,
π

where the infimum is taken over all strategies π with entropy rate r ≥ 0, and the quantization
error


(3)
D(q) (r|Y, d, s) := inf
d(Y, π(Y ))
s ,
π

the infimum being taken over all strategies π having quantization rate r ≥ 0. Often, all or
some of the parameters Y , d, s are clear from the context, and will therefore be omitted. The
quantization information constraint is more restrictive, so that the quantization error always
dominates the entropy coding error. Moreover, the coding error increases with the moment

under consideration.
702

Unless otherwise stated, we choose as original the fractional Wiener process Y = X. We are
mainly concerned with two particular choices for the distortion d. First we analyse the supremum
norm distortion that is d(f, g) = kf − gk[0,1] . In this setting we find:
Theorem 1.1. There exists a constant κ = κ(H) ∈ (0, ∞) such that for all s1 ∈ (0, ∞] and
s2 ∈ (0, ∞),
lim rH D(e) (r|s1 ) = lim rH D(q) (r|s2 ) = κ.
r→∞

r→∞

Remark 1.2. In the above theorem, general c`adl`ag functions are allowed as reconstructions.
Since the original process is continuous, it might seem more natural to use continuous functions
as approximations. The following argument shows that confining oneself to continuous approximants does not change the corresponding quantization and entropy quantity, when s ∈ [1, ∞).
Let π : C[0, 1] → D[0, 1] be an arbitrary strategy and let τn : D[0, 1] → C[0, 1] denote the linear
operator mapping f to its piecewise linear interpolation with supporting points 0, n1 , n2 . . . , 1.
Then

kX − τn ◦ π(X)k[0,1]
≤
kτn (X) − τn ◦ π(X)k[0,1]
+
kX − τn (X)k[0,1]
s
s

s


≤
kX − π(X)k[0,1]
+
kX − τn (X)k[0,1]

.
s

s

Note that the second term vanishes when n tends to infinity and that τn ◦ π satisfies the same
information constraint as π. The argument can be easily modified in order to show the statement
for s ∈ (0, 1).
Under Lp [0, 1]-norm distortion (p ≥ 1) that is d(f, g) = kf − gkLp [0,1] , we prove the following
analog to Theorem 1.1:
Theorem 1.3. For every p ≥ 1 there exists a constant κp = κp (H) ∈ (0, ∞) such that for all
s ∈ (0, ∞),
lim rH D(e) (r|s) = lim rH D(q) (r|s) = κp .
r→∞

r→∞

As in Remark 1.2, a simple convolution type argument shows that allowing Lp [0, 1]approximations yields the same coding errors as restricting oneself to C[0, 1]-approximations.
For ease of notation, the article is restricted to the analysis of 1-dimensional processes. However,
(1)

(d)
when replacing (Xt ) by a process (Xt , . . . , Xt ) consisting of d independent fractional Brownian motions, the proofs can be easily adapted, and one obtains analogous results. In particular,
it is possible to prove analogs of the above theorems for a multi-dimensional Brownian motion.
Let us summarize some of the known estimates for the constant κ in the case where X is standard
Brownian motion, i.e. H = 1/2.
• Under supremum-norm distortion, the relationship between the small ball function and
the quantization problem (see (8)) shows that
 π

κ ∈ √ ,π .
8
• Under Lp [0, 1]-norm distortion, κp may again be estimated via a connection to the small
ball function. Indeed, letting
Z ∞
o
nZ ∞
p 2
1
(ϕ′ (x))2 dx ,
|x| ϕ (x) dx + 2
λ1 = inf
−∞

−∞

703

where the infimum is taken over all weakly differentiable ϕ ∈ L2 (R) with unit norm, one
has
√
κp ∈ [c, 8 c]


√
λ1 (2+p)/2p
for c = 21/p p 2+p
.
In the case where p = 2, the constant κ2 is known explicitly: κ2 =

√

2
π

(see (17) and (4)).

The article is organized as follows. In Sections 2 to 5 we consider the approximation problems
under the supremum norm. We start in Section 2 by introducting a coding scheme which plays
an important role in the sequel. In Section 3, we use the construction of Section 2 and the
self similarity of X to establish a polynomial decay for D(e) (·|∞). In the following section,
the asymptotics of the quantization error are computed. The proof relies on a concentration
property for the entropies of “good” coding schemes (Proposition 4.4). In Section 5, we use the
equivalence of moments in the quantization problem to establish a lower bound for the entropy
coding problem. In the last section, we treat the case where the distortion is based on the
Lp [0, 1]-norm, i.e. d(f, g) = kf − gkLp [0,1] ; we introduce the distortion rate function and prove
Theorem 1.3 with the help of Shannon’s source coding theorem.
It is convenient to use the symbols ∼, . and ≈. We write f ∼ g iff lim fg = 1, while f . g
stands for lim sup fg ≤ 1. Finally, f ≈ g means
0 < lim inf

2

f
f
≤ lim sup < ∞ .
g
g

The coding scheme

This section is devoted to the construction of strategies π (n) : C[0, n] → D[0, n] which we will
need later in our discussion. The construction depends on three parameters: M ∈ N\{1}, d > 0
and a strategy π : C[0, 1] → D[0, 1].

The coding scheme is motivated as follows: Due to the self similarity of the fractional Wiener
process, coding X on [0, 1] with accuracy εn−H is as hard as coding X on the time interval
[0, n] with accuracy ε (see the argument at the end of the proof of Lemma 3.4). Intuitively, one
may decompose the coding scheme for (Xt )t∈[0,n] into two steps. First store information on the
values (Xj )j=1,...,n−1 and then approximate the paths X (j) = (Xj+t − Xj )t∈[0,1) by π(X (j) ) for
j = 0, . . . , n − 1. The parameter M governs the rate spent on coding the first part, and we shall
see that for ε small most rate is spent for coding the second part. As in Shannon’s source coding
theorem, the ergodicity of (X (j) )j∈N0 can be used to construct close to optimal codebooks when
n is large. This will be done in the proof of Theorem 4.1 to derive an upper bound for the
quantization error. Moreover, the coding scheme leads to a weak form of subadditivity which
we use to prove polynomial decay of D(e) (·|∞) (see Theorem 3.1 and Lemma 3.4).
(n)

We define the maps π (n) by induction. Let w ∈ C[0, ∞) and set (wt )t∈[0,1] := (wt+n − wn )t∈[0,1]
and w
ˆt := π(w(0) )(t) for t ∈ [0, 1). Assume that (w
ˆt )t∈[0,n) (n ∈ N) has already been defined.
Then we choose ξn to be the smallest number in {−d + 2kd/(M − 1) : k = 0, . . . , M − 1}
minimizing
|wn − (w
ˆn− + ξn )|,
704

and extend the definition of w
ˆ on [n, (n + 1)) by setting
w
ˆn+t := w
ˆn− + ξn + π(w(n) )(t),

t ∈ [0, 1).

Note that (w
ˆt )t∈[0,n) depends only upon (wt )t∈[0,n) , so that the above construction induces strategies
(n)
π (n) : C[0, n] → D[0, n], w 7→ (w
¯t )t∈[0,n] ,
(n)

where w
¯t

(n)

=w
ˆt for t ∈ [0, n) and w
¯n = w
ˆn− . Moreover, we can write
(n)

(w
¯t )t∈[0,n] = π (n) (w) = ϕn (π(w(0) ), . . . , π(w(n−1) ), ξ1 , . . . , ξn−1 )

(4)

for an appropriate measurable function ϕn : (D[0, n])n × Rn−1 → D[0, n].

The main motivation for this construction is the following property. If one has, for some (wt ) ∈
C[0, ∞) and n ∈ N,
M
kw − π (n) (w)k[0,n] ≤
d
M −1
and kw(n) − π(w(n) )k[0,1] ≤ d, then

|wn − (w
ˆn− + ξn )| ≤

d
,
M −1

whence,
(n)

kw − wk
ˆ [n,n+1) = kwn + wt

− (w
ˆn− + ξn + π(w(n) )(t))k[0,1)

≤ |wn − (w
ˆn− + ξn )| + kw(n) − π(w(n) )k[0,1)
M
d.
≤ d/(M − 1) + d =
M −1

In particular, if π : C[0, 1] → D[0, 1] satisfies


kX − π(X)k[0,1]
≤ d,
∞
then for any n ∈ N,



kX − π (n) (X)k[0,n]
≤
∞

3

M
d.
M −1

(5)

Polynomial decay of D(e) (r|∞)

The objective of this section is to prove the following theorem.
Theorem 3.1. There exists a constant κ = κ(H) ∈ (0, ∞) such that
lim rH D(e) (r|∞) = κ.

r→∞

(6)

Thereafter, κ = κ(H) will always denote the finite constant defined via equation (6). In order
to simplify notation, we abridge k · k = k · k[0,1] .
705

Remark 3.2. It was found in (4) (see Theorem 3.5.2) that for finite moments s ≥ 1 the entropy
coding error is related to the asymptotic behavior of the small ball function of the Gaussian
measure. In particular, for fractional Brownian motion, one obtains that
D(e) (r|s) ≈

1
,
rH

r → ∞.

In order to show that D(e) (r|∞) is of the order r−H , we still need to prove an appropriate upper
bound. We prove a stronger statement which will be useful later on.
Lemma 3.3. There exist strategies π (r) : C[0, 1] → D[0, 1], r ≥ 0, and probability weights
(r)
(pw )w∈im(π(r) ) such that for any s ≥ 1,


kX − π (r) (X)k
≤ 1
∞
rH

and

(r)

E[(− log pπ(r) (X) )s ]1/s ≈ r.

(7)

In particular, D(e) (r|∞) ≈ r−H .

The proof of the lemma is based on an asymptotic estimate for the mass concentration in
˜ 1 denote a fractional Brownian motion
randomly centered small balls, to be found in (9). Let X
˜
that is independent of X with L(X) = L(X1 ). Then, for any s ∈ [1, ∞), one has
˜ 1 k ≤ ε|X))s ]1/s ≈ − log P(kXk ≤ ε) ≈ ε−1/H
E[(− log P(kX − X

(8)

as ε ↓ 0 (see (9), Theorem 4.2 and Corollary 4.4).

Proof. For a given D[0, 1]-valued sequence (w
˜n )n∈N∪{∞} , we consider the following coding
(r)
strategy π (·|(w
˜n )): let
T (r) (w) := T (r) (w|(w
˜n )) := inf{n ∈ N : kw − w
˜n k ≤ 1/rH },
with the convention that the infimum of the empty set is ∞, and set
π (r) (w) := π (r) (w|(w
˜n )) := w
˜T (r) (w) .
Moreover, let (pn )n∈N denote the sequence of probability weights defined as
pn =

6 1
,
π 2 n2

n ∈ N,

and set p∞ := 0.
˜ n )n∈N∪{∞} denote independent FBM’s that are also independent of X, and analyze
Now we let (X
˜ n )). With T (r) := T (r) (X|(X
˜ n )) we obtain
the random coding strategies π (r) (·) := π (r) (·|(X
ˆ (r) := π (r) (X) = X
˜ (r) ,
X
T
and
E[(− log pT (r) )s ]1/s ≤ 2E[(log T (r) )s ]1/s + log
706

π2
.
6

(9)

˜1k ≤
Given X, the random time T (r) is geometrically distributed with parameter P(kX − X
H
1/r |X), and due to Lemma A.2 there exists a universal constant c1 = c1 (s) < ∞ for which
˜ 1 k ≤ 1/rH |X)].
E[(log T (r) )s |X]1/s ≤ c1 [1 + log E[T (r) |X]] = c1 [1 + log 1/P(kX − X
Consequently,

1/s
E[(log T (r) )s ]1/s = E E[(log T (r) )s |X]

˜ 1 k ≤ 1/rH |X))s ]1/s
≤ c1 E[(1 + log 1/P(kX − X
˜ 1 k ≤ 1/rH |X))s ]1/s ).
≤ c1 (1 + E[(− log P(kX − X

Due to (8), one has

(10)

˜ 1 k ≤ 1/rH |X))s ]1/s ≈ r,
E[(− log P(kX − X

so that (9) and (10) imply that E[(− log pT (r) )s ]1/s . c2 r for some appropriate constant c2 < ∞.
In particular, for any r ≥ 0, we can find a C[0, 1]-valued sequence (w
˜ (r) )n∈N of pairwise different
elements such that
E[(− log pT (r) (X|(w˜(r) )) )s ]1/s ≤ E[(− log pT (r) )s ]1/s . c2 r.
n

(r)

Now the strategies π (r) (·|(w
˜n )) with associated probability weights p
(7). Moreover, D(e) (r|∞) ≈ r−H follows since

(r)
H(π (r) (X|(w
˜n(r) ))) ≤ E − log p (r)
π

(r)

(X|(w
˜n ))

(r)
(r)

w
˜n

:= pn (n ∈ N) satisfy


.



Next we use the coding scheme of Section 2 to prove
Lemma 3.4. Let n ∈ N, r ≥ 0 and ∆r ≥ 1. Then
D(e) (n(r + ∆r)|∞) ≤ n−H

e∆r
D(e) (r|∞).
e∆r − 2

(11)

Proof. Fix ε > 0 and let π : C[0, 1] → D[0, 1] be a strategy satisfying


kX − π(X)k[0,1]
≤ (1 + ε)D(e) (r|∞) =: d
∞
and

H(π(X)) ≤ r.

Choose M := ⌊e∆r ⌋ and let π (n) be as in Section 2. Note that ∆r ≥ 1 guarantees that M ≥
e∆r − 1 ≥ e∆r /2, so that


kX − π (n) (X)k[0,n]
≤
∞

M
e∆r
d ≤ ∆r
(1 + ε)D(e) (r|∞).
M −1
e −2

707

(i)

We let (Xt )t∈[0,1] = (Xi+t − Xi )t∈[0,1] for i = 1, . . . , n, and (ξi )i=1,...,n−1 be as in Section 2 for
w = X. Observe that, due to the representation (4),
H(π (n) (X)) ≤ H(π(X (0) ), . . . , π(X (n−1) ), ξ1 , . . . , ξn−1 )

≤ H(π(X (0) )) + · · · + H(π(X (n−1) )) + H(ξ1 , . . . , ξn−1 )

≤ nr + log | range (ξ1 , . . . , ξn−1 )| ≤ nr + n log M

(12)

≤ n(r + ∆r).
Now let

αn : D[0, 1] → D[0, n], f 7→ αn (f )(s) = nH f (s/n)
and consider the strategy
π
˜ : C[0, 1] → D[0, 1], f 7→ αn−1 ◦ π (n) ◦ αn (f ).
Since αn (X) is again a fractional Brownian motion on [0, n], it follows that, a.s.
kX − π
˜ (X)k[0,1] = n−H kαn (X) − π (n) (αn (X))k[0,n] ≤ (1 + ε) n−H

e∆r
D(e) (r|∞).
e∆r − 2

Moreover,
H(˜
π (X)) = H(αn−1 ◦ π (n) (αn (X))) = H(π (n) (X)) ≤ r.
Since ε > 0 is arbitrary, the proof is complete.



Proof of Theorem 3.1. For r ≥ 0, ∆r ≥ 1 and n ∈ N, Lemma 3.4 yields
D(e) (n(r + ∆r)|∞) ≤

e∆r
1
D(e) (r|∞).
nH e∆r − 2

Now set κ := lim inf r→∞ rH D(e) (r|∞) which lies in (0, ∞) due to Lemma 3.3. Let ε ∈ (0, 1/2)
be arbitrary, and choose r0 , ∆r ≥ 1 such that

H
(e)

r0 D (r0 |∞) ≤ (1 + ε)κ,
∆r ≤ εr0 and

 −∆r
e
≤ ε.

Then

1
1
D(e) (r0 |∞)
H
n 1 − 2ε
1
1
≤
(1 + ε)1+H κ

H
1 − 2ε
(1 + ε)nr0

D(e) ((1 + ε)nr0 |∞) ≤

and we obtain that
lim sup (1 + ε)nr0
n→∞


H

D(e) ((1 + ε)nr0 |∞) ≤

708

(1 + ε)1+H
κ.
1 − 2ε

Let now r ≥ (1 + ε)r0 and introduce r¯ = r¯(r) = min{(1 + ε)nr0 : n ∈ N, r ≤ (1 + ε)nr0 } as well
as r = r(r) = max{(1 + ε)nr0 : n ∈ N, (1 + ε)nr0 ≤ r}. Using the monotonicity of D(e) (r|∞),
we conclude that
lim sup rH D(e) (r|∞) ≤ lim sup r¯H D(e) (r|∞)
r→∞

r→∞

≤ lim sup (r + (1 + ε)r0 )H D(e) (r|∞)
r→∞

≤

(1 + ε)1+H
κ.
1 − 2ε

Noticing that ε > 0 is arbitrary finishes the proof.

4



The quantization problem

Theorem 4.1. One has for any s ∈ (0, ∞),
D(q) (r|s) ∼ κ

1
,
rH

r → ∞.

Recall that a strategy π and probability weights (pw ) on the image of π intuitively correspond to
a coding scheme which maps an original symbol x onto a prefix-free representation for π(x) with
codelength of about − log2 pπ(x) . The proof of Theorem 4.1 relies on Proposition 4.4. There we
show that for good coding schemes − log2 pπ(X) is strongly concentrated around some typical
value when r is large. In order to prove the proposition we combine Lemma 3.3 with the following
lemma.
(r)

Lemma 4.2. There exist strategies (π (r) )r≥0 and probability weights (pw ) such that


kX − π (r) (X)k
≤ κ 1
∞
rH

and

(r)

− log pπ(r) (X) . r,

Proof. Let ε > 0 and choose r0 ≥ 2 such that
 r + 1 1/H
0

r0 − 1

≤1+

By Theorem 3.1,

D(e) ((1 + ε/2)r|∞) . κ

in probability.

ε
2

r0 − 1 1
r0 + 1 r H

In particular, there exists r1 ≥ r0 ∨ 2ε log(r0 + 1) and a map π : C[0, 1] → D[0, 1] such that


kX − π(X)k[0,1]
≤ κ r0 − 1 1 =: d and H(π(X)) ≤ (1 + ε/2)r1 .
∞
r0 r1H

For n ∈ N, let π (n) and ϕn be as in Section 2 for M = ⌈r0 ⌉, d and π. Then by (5)


kX − π (n) (X)k[0,n]
≤ κ (r0 − 1)M 1 ≤ κ 1 .
∞
r0 (M − 1) r1H
r1H
709

(13)

For w
ˆ (0) , . . . , w
ˆ (n−1) ∈ im(π) and k1 , . . . , kn−1 ∈ {−d +
(n)
pϕ (wˆ(0) ,...,wˆ(n−1) ,k ,...,k )
n
1
n−1

=

1
M n−1

2kd
M −1
n−1
Y

: k = 0, . . . , M − 1}, let

P(π(X) = w
ˆ (i) ).

i=0

(n)

The (pw ) define probability weights on the image of ϕn . Moreover,
− log p

(n)
ˆ t )t∈[0,n]
(X

= (n − 1) log M −

n−1
X

log pπ(X (i) )

i=0

and the ergodic theorem implies
lim −

n→∞

1
(n)
log p ˆ
= log M + H(π(X)),
(Xt )t∈[0,n]
n

a.s.

Note that log M + H(π(X)) ≤ (1 + ε)r1 .

Just as in the proof of Lemma 3.4, we use the self similarity of X to translate the strategy π (n)
into a strategy for encoding (Xt )t∈[0,1] . For n ∈ N, let
αn : D[0, 1] → D[0, n], f 7→ (αn f )(t) = nH f (t/n)
(n)

(n)

and consider p˜w := pαn (w) and π
˜ (n) (w) := αn−1 ◦ π (n) ◦ αn (w). Then
(n)

(n)

− log p˜π˜ (n) (X) = − log pπ(n) (α

n (X))

. (1 + ε)nr1 ,

in probability

and by (13)




kX − π
˜ (n) (X)k[0,1]
∞ =
kαn−1 (αn (X) − π (n) (αn (X)))k[0,1]
∞

1
= H
kαn (X) − π (n) (αn (X))k[0,n]
∞
n

1
1
.
= H
kX − π (n) (X)k[0,n]
∞ ≤ κ
n
(nr1 )H

By choosing π
¯ (r) = π
˜ (n) and (¯
p(r) ) = (˜
p(n) ) for r ∈ ((n − 1)r1 , nr1 ], one obtains a coding scheme
satisfying


1
kX − π
¯ (r) (X)k
∞ ≤ κ H
r
and
(r)
− log p¯π¯ (r) (X) . (1 + ε)r, in probability,
so that the assertion follows by a diagonalization argument.



Remark 4.3. In the above proof, we have constructed a high resolution coding scheme based
on a strategy π : C[0, 1] → D[0, 1], using the identity π
˜ (n) = αn−1 ◦ π (n) ◦ αn . This coding scheme
leads to a coding error which is at most

M

kX − π(X)k[0,1]
n−H .
∞
M −1
710

(14)

Moreover, the ergodic theorem implies that, for large n, π
˜ (n) (X) lies with probability almost one
(n)
in the typical set {w ∈ D[0, 1] : − log p˜w ≤ n(H(π(X)) + log M + ε)}, where ε > 0 is arbitrarily
small. This set is of size exp{n(H(π(X)) + log M + ε)}, and will serve as a close to optimal high
resolution codebook. It remains to control the case where π
˜ (n) (X) is not in the typical set. We
will do this in the proof of Theorem 4.1 at the end of this section (see (19)).
(r)

Proposition 4.4. For s ≥ 1 there exist strategies (π (r) )r≥0 and probability weights (pw ) such
that
(r)



kX − π (r) (X))k
≤ κ 1
∞
rH

and

lim

E[(− log pπ(r) (X) )s ]1/s
r

r→∞

= 1.

(15)

In addition, for any ε > 0 one has

1 
lim sup P − log pπ(X) ≤ (1 − ε)r, kX − π(X)k ≤ κ H = 0,
r→∞ π,(p )
r
w

(16)

where the supremum is taken over all strategies π : C[0, 1] → D[0, 1] and over all sequences of
probability weights (pw ).
(r)

(r,1)

Proof. Fix s > 1 and let for each R ≥ 0, π1 be a strategy and (pw ) be a sequence of
(r)
(r,2)
probability weights as in Lemma 4.2. Moreover, let π2 and (pw ) be as in Lemma 3.3 for the
(r)
(r,2)
(r)
(r,1)
and κ2 (w) := − log p (r) , and
moment 2s. We consider the maps κ1 (w) := − log p (r)
π (r) (w) :=

( (r)
π1 (w)
(r)

π2 (w)

(r)

if κ1 (w) ≤ (1 + δ)r,
otherwise,

(r)

for some fixed δ > 0. Then one obtains, for pw =
(r)
κ1 (w)

π2 (w)

π1 (w)

set

1 (r,1)
2 (pw

(r,2)

+ pw

≤ (1 + δ)r},

) and Tr := {w ∈ C[0, 1] :

(r)

(r)

(r)

E[(− log 2pπ(r) (X) )s ]1/s ≤ E[1Tr (X)κ1 (X)s ]1/s + E[1Trc (X)κ2 (X)s ]1/s
(r)

≤ (1 + δ)r + P(X ∈ Trc )1/2s E[κ2 (X)2s ]1/2s .
(r)

The definitions of π1
Consequently,

(r)

and π2

(r)

imply that limr→∞ P(X ∈ Trc ) = 0 and E[κ2 (X)2s ]1/2s ≈ r.
(r)

E[(− log pπ(r) (X) )s ]1/s . (1 + δ)r.
Since δ > 0 can be chosen arbitrarily small, a diagonalization procedure leads to strategies π
˜ (r)
(r)
and probability weights (˜
pw ) with


1
kX − π
˜ (r) (X)k[0,1]
∞ ≤ κ H and E[(− log p˜π˜ (r) (X) )s ]1/s . r.
r

Now the first assertion follows from (16).

(r)

It remains to show that for arbitrary strategies π
¯ (r) , r ≥ 0, and probability weights (¯
pw ):

1 
(r)
¯ (r) (X)k ≤ κ H = 0.
lim P − log p¯π¯ (r) (X) ≤ (1 − ε)r, kX − π
(17)
r→∞
r
711

Without loss of generality, we can assume that


1
kX − π
¯ (r) (X)k[0,1]
∞ ≤ κ H .
r

(18)

Otherwise we modify the map π
¯ (r) for all w ∈ C[0, 1] with kw − π
¯ (r) (w)k > κ r−H in such a way
that (18) be valid. Hereby the probability in (17) increases and it suffices to prove the statement
for the modified strategy. Let us consider
(
(r)
(r)
π
¯ (r) (w) if p¯π¯ (r) (w) ≥ p˜π˜ (r) (w)
(r)
π (w) =
π
˜ (r) (w) else.
p(r) + p˜(r) ) satisfy
Then the probability weights p(r) := 12 (¯
(r)

(r)

E[(− log 2pπ(X) )s ]1/s ≤ E[(− log p˜π˜ (X) )s ]1/s . r.
Recall that



kX − π (r) (X)k[0,1]
≤ κ 1 ,
∞
rH
(r)

hence by Theorem 3.1, one has E[− log pπ(r) (X) ] ≥ H(π (r) (X)) & r. Now the equivalence of
moments (see Lemma A.1) implies that
(r)

− log pπ(r) (X) ∼ r,
and

(r)

(r)

in probability,

− log p¯π¯ (r) (X) ≥ − log 2pπ(r) (X) & r,

in probability,

which gives (17).



Proof of Theorem 4.1. We start by proving the lower bound. Fix s > 0, let Cr , r ≥ 0,
denote arbitrary codebooks of size er , and let π (r) : C[0, 1] → Cr denote arbitrary strategies.
(r)
(r)
Moreover, let (pw ) be the sequence of probability weights defined as pw = 1/|Cr |, w ∈ Cr . Then
(r)
− log pπ(r) (X) ≤ r a.s., and the above lemma implies that for any ε ∈ (0, 1),

Therefore,


(1 − ε)H 
lim P kX − π (r) (X)k ≤ κ
= 0.
r→∞
rH

E[kX − π (r) (X)ks ]1/s ≥ κ
which proves the lower bound.

(1 − ε)H 1/s
(1 − ε)H
(1 − ε)H 
(r)
P
kX
−
π
(X)k
≥
κ
,
∼
κ
rH
rH
rH

It remains to show that D(q) (r|s) . κ/rH . By Lemma 4.2, there exist strategies π (r) and
(r)
probability weights (pw ) such that


kX − π (r) (X)k
≤ κ 1 and − log p (r)
π (X) . r,
∞
rH
712

in probability.

Furthermore, due to Theorem 4.1 in (8), there exist codebooks C¯r of size er with
E[ min kX − wk
ˆ 2s ]1/2s ≈
w∈
ˆ C¯r

1
.
rH

(r)
We consider the codebook Cr := C¯r ∪ {w
ˆ : − log pwˆ ≤ (1 + ε/2)r}. Clearly, Cr contains at most
er + e(1+ε/2)r elements. Moreover,

E[ min kX − wk
ˆ s ]1/s ≤ E[1Cr (π (r) (X)) (κ
w∈C
ˆ r

1 s 1/s
) ]
rH

+ E[1Crc (π (r) (X)) min kX − wk
ˆ s ]1/s
w∈
ˆ C¯r

≤κ

(19)

1
+ P(π (r) (X) 6∈ Cr )1/2s E[ min kX − wk
ˆ 2s ]1/2s .
rH
w∈
ˆ C¯r

Since limr→∞ P(π (r) (X) 6∈ Cr ) = 0 and the succeeding expectation is of order O(1/rH ), the
second summand is of order o(1/rH ). Therefore, for r ≥ 2/ε
D(q) ((1 + ε)r|s) ≤ E[ min kX − wk
ˆ s ]1/s . κ
w∈C
ˆ r

1
.
rH

By switching from r to r˜ = (1 + ε)r, we obtain
D(q) (˜
r|s) . κ (1 + ε)H

1
.
r˜H

Since ε > 0 was arbitrary, the proof is complete.

5



Implications of the equivalence of moments

In this section we complement Theorem 4.1 by
Theorem 5.1. For arbitrary s ∈ (0, ∞], one has
D(e) (r|s) ∼ κ

1
.
rH

The proof of this theorem is based on the following general principle: if the asymptotic quantization error coincides for two different moments s1 < s2 , then all moments s ≤ s2 lead to the
same asymptotic quantization error and the entropy coding problem coincides with the quantization problem for all moments s ≤ s2 .
ˆ denoting arbitrary measurable spaces
Let us prove this relationship in a general setting. E and E
ˆ
and d : E × E → [0, ∞) a measurable function, the quantization error for a general E-valued
r.v. X under the distortion d is defined as
D(q) (r|s) = inf E[min d(X, x
ˆ)s ]1/s ,
C⊂E

x
ˆ∈C

ˆ with |C| ≤ er . In order to simplify
where the infimum is taken over all codebooks C ⊂ E
notations, we abridge
ˆ
d(x, A) = inf d(x, y),
x ∈ E, A ⊂ E.
y∈A

713

Analogously, we denote the entropy coding error by
ˆ s ]1/s ,
D(e) (r|s) = inf E[d(X, X)
ˆ
X

ˆ
ˆ with H(X)
ˆ ≤ r.
where the infimum is taken over all discrete E-valued
r.v. X

Then Theorem 5.1 is a consequence of Theorem 4.1 and the following theorem.
Theorem 5.2. Assume that f : [0, ∞) → R+ is a decreasing, convex function satisfying
+

−r ∂∂r f (r)
lim sup
< ∞,
f (r)
r→∞

(20)

and suppose that, for some 0 < s1 < s2 ,
D(q) (r + log 2|s1 ) ∼ D(q) (r|s2 ) & f (r).
Then for any s > 0,
D(e) (r|s) & f (r).
We need two technical lemmas.
Lemma 5.3. Let 0 < s1 < s2 and f : [0, ∞) → R+ . If
D(q) (r + log 2|s1 ) ∼ D(q) (r|s2 ) ∼ f (r),
then for any ε > 0,
lim sup P(d(X, C) ≤ (1 − ε)f (r)) = 0.

r→∞ C⊂E:
|C|≤er

Proof. For r ≥ 0, let Cr∗ denote codebooks of size er with
E[d(X, Cr∗ )s2 ]1/s2 ∼ f (r).

(21)

Now let Cr denote arbitrary codebooks of size er , and consider the codebooks C¯r := Cr∗ ∪ Cr .
Using (21) and the inequality s1 ≤ s2 , it follows that
f (r) & E[d(X, C¯r )s2 ]1/s2 ≥ E[d(X, C¯r )s1 ]1/s1 ≥ D(q) (r + log 2|s1 ) ∼ f (r).
Thus the s1 -th and the s2 -th moment coincide asymptotically and it follows by Lemma A.1 that
d(X, C¯r ) ∼ f (r),

in probability,

d(X, Cr ) & f (r),

in probability.

so that in particular,


714

Lemma 5.4. Assume that f : [0, ∞) → R+ is a decreasing, convex function satisfying (20) and
lim sup P(d(X, C) ≤ f (r)) = 0.

r→∞

ˆ
C⊂E:
|C|≤er

Then for any s > 0,
D(e) (r|s) & f (r).
Proof. The result is a consequence of the technical Lemma A.3. Consider the family F
consisting of all random vectors
ˆ s , − log p ˆ ),
(A, B) = (d(X, X)
X
ˆ is an arbitrary discrete E-valued
ˆ
where X
r.v. and (pw ) is an arbitrary sequence of probability
˜
ˆ
ˆ and (pw ) and an
weights on the range of X. Let f (r) = f (r)s , r ≥ 0. Then for any choice of X
r
ˆ : − log pw ≤ r} contains at most e elements. Consequently,
arbitrary r ≥ 0, the set C := {w ∈ E
ˆ ≤ f (r), X
ˆ ∈ C) ≤ P(d(X, C) ≤ f (r)).
ˆ s ≤ f˜(r), − log p ˆ ≤ r) = P(d(X, X)
P(d(X, X)
X

ˆ
By assumption the right hand side converges to 0 as r → ∞, independently of the choice of X
and (pw ). Since f˜ satisfies condition (27), Lemma A.3 implies that
D(e) (r|s) =

inf

ˆ
ˆ
X:H(
X)≤r

ˆ s ]1/s = inf E[A]1/s & f˜(r)1/s = f (r),
E[d(X, X)
A∈Fr

where Fr = {A : (A, B) ∈ F, EB ≤ r}.



Theorem 5.2 is now an immediate consequence of Lemma 5.3 and Lemma 5.4.

6

Coding with respect to the Lp [0, 1]-norm distortion

In this section, we consider the coding problem for the fractional Brownian motion X under
Lp [0, 1]-norm distortion for some fixed p ∈ [1, ∞). In order to treat this approximation problem,
we need to introduce Shannon’s distortion rate function. It is defined as


ˆ Lp [0,1]
,
D(r|s) = inf
kX − Xk
s

ˆ satisfying the mutual information
where the infimum is taken over all D[0, 1]-valued r.v.’s X
ˆ ≤ r. Here and elsewhere I denotes the Shannon mutual information, defined
constraint I(X; X)
as
(R
dP X
ˆ
log dPXX,
dPX,Xˆ if PX,Xˆ ≪ PX ⊗ PXˆ
⊗PX
ˆ
ˆ
I(X; X) =
∞
else.

The objective of this section is to prove

Theorem 6.1. The following limit exists
κp = κp (H) = lim rH D(r|p) ∈ (0, ∞),
r→∞

(22)

and for any s > 0, one has
D(q) (r|s) ∼ D(e) (r|s) ∼ κp
715

1
.
rH

(23)

We will first prove that statement (23) is valid for
κp := lim inf rH D(r|p).
r→∞

Since D(r|p) is dominated by D(q) (r|p), the existence of the limit in (22) then follows immediately. Due to Theorem 1.2 in (5), the distortion rate function D(·|p) has the same weak
asymptotics as D(q) (·|p). In particular, D(r|p) ≈ r−H and κp lies in (0, ∞).
We proceed as follows: decomposing X into the two processes
X (1) = (Xt − X⌊t⌋ )t≥0

and X (2) = (X⌊t⌋ )t≥0 ,

we consider the coding problem for X (1) and X (2) in Lp [0, n] (n ∈ N being large). We control
the coding complexity of the first term via Shannon’s source coding theorem (SCT) and use a
limit argument in order to show that the coding complexity of X (2) is asymptotically negligible.
We recall the SCT in a form which is appropriate for our discussion; for n ∈ N, let
1/p
Z 1
|f (t) − g(t)|p dt
dp (f, g) =
0

and
dn,p (f, g) =

Z

0

(f, g)p ,

n

|f (t) − g(t)|p

dt 1/p
.
n

Then dn,p
n ∈ N, is a single letter distortion measure, when interpreting the function
f |[0,n) as the concatenation of the “letters” f (0) , . . . , f (n−1) , where f (i) = (f (i + t))t∈[0,1) . Analogously, the process X (1) corresponds to the letters X (1,i) := (Xi+t )t∈[0,1) , i ∈ N0 . Since
(X (1,i) )i∈N0 is an ergodic stationary C[0, 1)-valued process, the SCT implies that for fixed r ≥ 0
and ε > 0 there exist codebooks Cn ⊂ D[0, n], n ∈ N, with at most exp{(1 + ε)nr} elements such
that
lim P(dn,p (X (1) , Cn )p ≤ (1 + ε)D(r|p)p ) = 1.

n→∞

(24)

The statement is an immediate consequence of the asymptotic equipartition property as stated
in (3) (Theorem 1) (see also (1) and (3)).
First we prove a lemma which will later be used to control the coding complexity of X (2) .
Lemma
6.2. Let (Zi )i∈N be an ergodic stationary sequence of real-valued r.v.’s and let Sn =
Pn
n
Z
,
i=1 i n ∈ N0 . For ε > 0 there exist codebooks Cn ⊂ 2εZ of size exp{2nE[log(|Z1 |/2ε + 2)] +
2nc} satisfying


n ) ≤ ε = 1,
lim P min kS1n − sˆkl∞
n→∞

S1n

where
on Rn .

sˆ∈Cn

n denotes the maximum norm
denotes (Si )i=1,...,n , c is a universal constant and k · kl∞

Proof. Let c > 0 be such that (pn )n∈Z defined through
pn = e−c

1
(|n| + 1)2

716

is a sequence of probability weights, and let


1
(n)
Cn = sˆn1 ∈ 2εZn : − log psˆn ≤ 2E[log(|Z1 |/2ε + 2)] + 2c ,
1
n

where

(n)
psˆn
1

= psˆ1

n
Y
i=2

(n)

sˆn1 ∈ 2εZn .

p(ˆsi −ˆsi−1 )/2ε ,

Since (psˆn ) defines a sequence of probability weights on 2εZn the set Cn satisfies the required
1
size constraint. Let Sˆ1n denote a best approximation for S1n in the set 2εZ. Then always
n ≤ ε. Note that
kS1n − Sˆ1n kl∞
(n)
S1

− log p ˆn ≤ 2

n
X
i=1

log(|Zi |/2ε + 2) + nc

so that the ergodic theorem implies that limn→∞ P(Sˆ1n ∈ Cn ) = 1 which implies the assertion. 
We now use the SCT combined with the previous lemma to construct codebooks that guarantee
almost optimal reconstructions with a high probability.
Lemma 6.3. For any ε > 0 there exist codebooks Cr , r ≥ 0, of size er such that
lim P(dp (X, Cr ) ≤ (1 + ε)κp r−H ) = 1.

r→∞

Proof. Let ε > 0 be arbitrary and c be as in Lemma 6.2. We fix r0 ≥

4εκp
1/H
E|X1 |

εκp r−H ≥ e−εr+c+log E|X1 | and D(r0 |p) ≤ (1 + ε)κp r0−H .

such that
(25)

We decompose X into the two processes
(1)

Xt

= Xt − X⌊t⌋

(2)

and Xt

= X⌊t⌋ .

(1)

Due to the SCT (24), there exist codebooks Cn ⊂ D[0, n] of size exp{(1 + ε)nr0 } satisfying
lim P(dn,p (X (1) , Cn(1) )p ≤ (1 + 2ε)p κpp r0−pH ) = 1.

n→∞

We apply Lemma 6.2 for ε′ := εκp r0−H . Note that
E log
Since r0H ≥

4εκp
E|X1 | ,

 |X |

 E|X |

1
1
+
2
+
c
≤
log
+
2
+c
2ε′
2ε′

it follows that
E log

E|X1 |
2ε′

=

r0H E|X1 |
2εκp

≥ 2, so that


 E|X | 
 |X |
1
1
+
2
+
c
≤
log
+c
2ε′
ε′
= − log(εκp r0−H ) + c + log E|X1 | ≤ εr,
717

(2)

due to (25). Hence, there exist codebooks Cn ⊂ D[0, n] of size exp{εnr0 } with

1 
lim P dn,p (X (2) , Cn(2) ) ≤ εκp H = 1.
n→∞
r0

(1)
(2)
(1)
(2)
Let now C˜n := Cn + Cn denote the Minkowski sum of the sets Cn and Cn . Then |C˜n | ≤
exp{(1 + 2ε)nr0 }, and one has

P(dn,p (X, C˜n ) ≤ (1 + 3ε)κp r0−H ) ≥ P(dn,p (X (1) , Cn(1) ) ≤ (1 + 2ε)κp r0−H and
dn,p (X (2) , Cn(2) ) ≤ εκp r0−H ) → 1.

Consider the isometric isomorphism
βn : Lp [0, 1] → (Lp [0, n], dn,p ), f 7→ f (nt),
and the codebooks Cn ⊂ D[0, 1] given by

Cn = {n−H βn−1 (w)
ˆ :w
ˆ ∈ C˜n }

˜ (n) = n−H βn−1 (X) is a fractional Brownian motion and one has
Then X
˜ (n) , Cn ) = dn,p (βn (X
˜ (n) ), βn (Cn )) = n−H dn,p (X, C˜n ).
dp (X
Hence, the codebooks Cn are of size exp{(1 + 2ε)nr0 } and satisfy

P(dp (X, Cn ) ≤ (1 + 3ε)κp (nr0 )−H )) = P(dn,p (X, C˜n ) ≤ (1 + 3ε)κp r0−H ) → 0

as n → ∞. Now the general statement follows by an interpolation argument similar to that used

at the end of the proof of Theorem 3.1.
(1)

Proof of Theorem 6.1. Let s ≥ 1 be arbitrary, let Cr be as in the above lemma for some
(2)
fixed ε > 0. Moreover, we let Cr denote codebooks of size er with
E[dp (X, Cr(2) )2s ]1/(2s) ≈
(1)

(2)

Then the codebooks Cr := Cr ∪ Cr
proof of Theorem 4.1 (see (19)),

1
.
rH

contain at most 2er elements and satisfy, in analogy to the

E[dp (X, Cr )s ]1/s . (1 + ε)κp

1
,
rH

r → ∞.

Since ε > 0 is arbitrary, it follows that
1
.
rH
For s ≥ p the quantization error is greater than the distortion rate function D(r|p), so that the
former inequality extends to
lim rH D(q) (r|s) = κp .
D(q) (r|s) . κp

r→∞

In particular, we obtain the asymptotic equivalence of all moments s1 , s2 greater or equal to p.
Next, an application of Theorem 5.2 with d(f, g) = dp (f, g)s implies that for any s > 0,
D(e) (r|s) & κp
which establishes the assertion.

1
,
rH


718

Appendix
Lemma A.1. For r ≥ 0, let Ar denote [0, ∞)-valued r.v.’s. If one has, for 0 < s1 < s2 and
some function f : [0, ∞) → R+ ,
E[Asr1 ]1/s1 ∼ E[Asr2 ]1/s2 ∼ f (r),

(26)

then
Ar ∼ f (r), in probability.
Proof. Consider

A˜r := Asr1 /E[Asr1 ],

and s˜2 = s2 /s1 . Then (26) implies that
E[A˜sr˜2 ]1/˜s2 ∼ E[A˜r ] = 1
so that by a classical result limr→∞ A˜r = 1 in Ls˜2 (P) or, equivalently, limr→∞ Ar /f (r) = 1 in
Ls2 (P). This immediately implies the result.

Lemma A.2. Let s ≥ 1. There exists a constant c = c(s) < ∞ such that for all [1, ∞)-valued
r.v.’s Z one has
E[(log Z)s ]1/s ≤ c [1 + log E[Z]].
Proof. Using elementary analysis, there exists a positive constant c1 = c1 (s) < ∞ such that
ψ(x) := (log x)s + c1 log x, x ∈ [1, ∞), is concave. For any [1, ∞)-valued r.v. Z, Jensen’s
inequality then yields
E[(log Z)s ]1/s ≤ E[ψ(Z)]1/s ≤ ψ(E[Z])1/s
1/s

≤ log E[Z] + c1 (log E[Z])1/s ≤ c [1 + log E[Z]],

where c = c(s) < ∞ is an appropriate universal constant.



Lemma A.3. Let f : [0, ∞) → R+ be a decreasing, convex function satisfying limr→∞ f (r) = 0
and
+
−r ∂∂r f (r)
lim sup
< ∞,
(27)
f (r)
r→∞
and F be a family of [0, ∞]2 -valued random variables for which
lim

sup P(A ≤ f (r), B ≤ r) = 0.

r→∞ (A,B)∈F

Then the sets of random variables Fr defined for r ≥ 0 through
Fr := {A : (A, B) ∈ F, EB ≤ r}
satisfy
inf EA & f (r)

A∈Fr

as r → ∞.
719

(28)

+

Proof. Fix R > 0, positive integers I and N , and define λ := − ∂∂r f (R),
ri :=

i+N
R,
N

i = −N, −N + 1, . . . .

For (A, B) ∈ F with EB ≤ r, we define
TA,B := {∄i ∈ {−N + 1, . . . , I} such that A ≤ f (ri ) and B ≤ ri }.
Then we have


E A + λB ≥
≥
=

I−1
X

i=−N
I−1
X

i=−N
I−1
X

i=−N

≥

I−1
X

i=−N



E 1TA,B 1[ri ,ri+1 ) (B)(A + λri )


E 1TA,B 1[ri ,ri+1 ) (B)(f (ri+1 ) + λri )
h
R i
E 1TA,B 1[ri ,ri+1 ) (B)(f (ri+1 ) + λri+1 − λ )
N
h
R i
E 1TA,B 1[ri ,ri+1 ) (B)(f (R) + λR − λ ) ,
N

where the last inequality follows from the fact that
f (R) + λR = inf [f (r) + λr]
r≥0

by the definition of λ and the convexity of f . Now, fix ε > 0 and pick N ≥ 1/ε, I ≥ 2N/ε and
R0 so large that
ε
P(TA,B ) ≥ 1 − for all R ≥ R0 and all (A, B) ∈ FR .
2
Using Chebychev’s inequality, we then obtain for R ≥ R0 ,



I
c
E[A + λB] ≥ (1 − ε)(f (R) + λR) 1 − P (T ) − P B ≥ R
N

ε ε
≥ (1 − ε)(f (R) + λR) 1 − −
.
2 2
Hence,

λR + EA ≥ (1 − ε)2 (f (R) + λR)
and therefore


EA ≥ (1 − ε)2 f (R) + λR (1 − ε)2 − 1 .

Using the definition of λ and (27), as well as the fact that ε > 0 is arbitrary, the conclusion
follows.


720

References
[1] T. M. Cover and J. A. Thomas. Elements of information theory. Wiley Series in Telecommunications. New York: John Wiley & Sons, Inc., 1991. MR1122806
[2] J. Creutzig. Approximation of Gaussian random vectors in Banach spaces. Ph.D. Dissertation, Friedrich-Schiller-Universit¨
at Jena, 2002.
[3] A. Dembo and I. Kontoyiannis. Source coding, large deviations, and approximate pattern
matching. IEEE Trans. Inform. Theory, 48(6):1590–1615, 2002. Special issue on Shannon
theory: perspective, trends, and applications. MR1909475
[4] S. Dereich. High resolution coding of stochastic processes and small ball probabilities. Ph.D.
Dissertation, TU Berlin,
URL: http://edocs.tu-berlin.de/diss/2003/dereich steffen.htm, 2003.
[5] S. Dereich. Asymptotic behavior of the distortion-rate function for Gaussian processes in
Banach spaces. Bull. Sci. Math., 129(10):791–803, 2005. MR2178943
[6] S. Dereich. The coding complexity of diffusion processes under Lp [0, 1]-norm distortion.
Preprint, 2006.
[7] S. Dereich. The coding complexity of diffusion processes under supremum norm distortion.
Preprint, 2006.
[8] S. Dereich, F. Fehringer, A. Matoussi, and M. Scheutzow. On the link between small ball
probabilities and the quantization problem for Gaussian measures on Banach spaces. J.
Theoret. Probab., 16(1):249–265, 2003. MR1956830
[9] S. Dereich and A. Lifshits. Probabilities of randomly centered small balls and quantization
in Banach spaces. Ann. Probab., 33(4):1397–1421, 2005. MR2150193
[10] S. Dereich, T. M¨
uller-Gronbach, and K. Ritter. Infinite-dimensional quadrature and quantization. Preprint, arXiv:math.PR/0601240, 2006.
[11] S. Graf and H. Luschgy. Foundations of quantization for probability distributions. Lecture
Notes in Mathematics 1730, Berlin: Springer, 2000. MR1764176
[12] S. Graf, H. Luschgy, and G. Pag`es. Optimal quantizers for Radon random vectors in a
Banach space. Preprint, 2005.
[13] R. M. Gray and D. L. Neuhoff. Quantization. IEEE Trans. Inf. Theory, 44(6):2325–2383,
1998. MR1658787
[14] J. Kuelbs and W. V. Li. Metric entropy and the small ball problem for Gaussian measures.
J. Funct. Anal., 116(1):133–157, 1993. MR1237989
[15] W. V. Li and Q.-M. Shao. Gaussian processes: inequalities, small ball probabilities and
applications. In Stochastic processes: theory and methods, volume 19 of Handbook of Statist.,
pages 533–597. North-Holland, Amsterdam, 2001. MR1861734

721

[16] H. Luschgy and G. Pag`es. Functional quantization of Gaussian processes. J. Funct. Anal.,
196(2):486–531, 2002. MR1943099
[17] H. Luschgy and G. Pag`es. Sharp asymptotics of the functional quantization problem for
Gaussian processes. Ann. Probab., 32(2):1574–1599, 2004. MR2060310
[18] G. Pag`es and J. Printems. Functional quantization for pricing derivatives. Preprint, Universit´e de Paris VI, LPMA no. 930, 2004.

722